This issue of Econometric Theory marks 20 full years of publication of the journal ET and the beginning of our third decade. To celebrate the occasion, I invited a group of econometricians, including all the present and past co-editors of ET, to reflect on the future of econometrics. To focus discussion, I selected a theme that relates to a recent trend in the practical use of econometric methodology, one that seems destined to play an increasing role in the future of the subject—automated modeling and inference. Accordingly, this issue contains contributions on various aspects of the theme of automation, introducing the notion of automated discovery, analyzing the validity of selection procedures, and discussing the practical utility of these automated methods in econometric modeling exercises.

Our subject is the notion of automated discovery in econometrics. Advances in computer power, electronic communication, and data collection processes have all changed the way econometrics is conducted. These advances have helped to elevate the status of empirical research within the economics profession in recent years, and they now open up new possibilities for empirical econometric practice. Of particular significance is the ability to build econometric models in an automated way according to an algorithm of decision rules that allow for (what we call here) heteroskedastic and autocorrelation robust (HAR) inference. Computerized search algorithms may be implemented to seek out suitable models, thousands of regressions and model evaluations may be performed in seconds, statistical inference may be automated according to the properties of the data, and policy decisions can be made and adjusted in real time with the arrival of new data. We discuss some aspects and implications of these exciting, emergent trends in econometrics.The first version of this paper was written in April 2004 for the 20th Anniversary Issue of Econometric Theory. Helpful comments by the co-editor, Oliver Linton, Benno Pötscher, Brendan Beare, and two referees on the first draft are gratefully acknowledged.

Model selection has an important impact on subsequent inference. Ignoring the model selection step leads to invalid inference. We discuss some intricate aspects of data-driven model selection that do not seem to have been widely appreciated in the literature. We debunk some myths about model selection, in particular the myth that consistent model selection has no effect on subsequent inference asymptotically. We also discuss an “impossibility” result regarding the estimation of the finite-sample distribution of post-model-selection estimators.

Standard econometric model selection methods are based on four conceptual errors: parametric vision, the assumption of a true data generating process, evaluation based on fit, and ignoring the impact of model uncertainty on inference. Instead, econometric model selection methods should be based on a semiparametric vision, models should be viewed as approximations, models should be evaluated based on their purpose, and model uncertainty should be incorporated into inference methods. These problems have been examined individually but not jointly, and my view is that future research into econometric model selection should attempt to address all four issues.This research was supported by the National Science Foundation. I thank Peter Phillips and a referee for helpful comments that greatly improved the arguments and exposition.

When Stephen Perez and I first began our Monte Carlo studies of the efficacy of general-to-specific search methodologies in 1995, we were keenly aware of our limited ability to capture the tacit knowledge of the skilled time-series econometrician operating in the London School of Economics (LSE) tradition (Hoover and Perez, 1999a, 1999b). Econometrics, we believed, was an art, and our algorithm was not intended to replace the artist. David Hendry and Hans-Martin Krolzig's subsequent development of PcGets did not, in fact, eliminate the art of econometrics. Power tools did not eliminate the art of the cabinetmaker but changed where his value added lay and—importantly—made new things possible. PcGets is likewise a new, powerful tool, useful in the hands of a skilled craftsman.I thank Peter Phillips, Selva Demiralp, Stephen Perez, and an anonymous referee for helpful comments on an earlier draft.

This note discusses the (dis-)similarities between automated inference and computer-aided decisions, at the interface of econometrics and economics. It is argued that computer-aided decisions are best suited for scientific communication. For the future, the topic of learning is singled out as one of the most promising areas of integration of econometric techniques and economics.I thank Peter Phillips and an anonymous referee for useful remarks on a previous version of this comment. The usual disclaimer applies.

Infinite order vector autoregressive (VAR) models have been used in a number of applications ranging from spectral density estimation, impulse response analysis, and tests for cointegration and unit roots, to forecasting. For estimation of such models it is necessary to approximate the infinite order lag structure by finite order VARs. In practice, the order of approximation is often selected by information criteria or by general-to-specific specification tests. Unlike in the finite order VAR case these selection rules are not consistent in the usual sense, and the asymptotic properties of parameter estimates of the infinite order VAR do not follow as easily as in the finite order case. In this paper it is shown that the parameter estimates of the infinite order VAR are asymptotically normal with zero mean when the model is approximated by a finite order VAR with a data dependent lag length. The requirement for the result to hold is that the selected lag length satisfies certain rate conditions with probability tending to one. Two examples of selection rules satisfying these requirements are discussed. Uniform rates of convergence for the parameters of the infinite order VAR are also established.Very helpful comments by the editor and two referees led to a substantial improvement of the manuscript. I am particularly indebted to one of the referees for pointing out an error in the proofs. All remaining errors are my own. Financial support from NSF grant SES−0095132 is gratefully acknowledged.

A simple regression approach to HAC and LRV estimation is suggested. The method exploits the fact that the quantities of interest relate to only one point of the spectrum (the origin). The new estimator is simply the explained sum of squares in a linear regression whose regressors are a set of trend basis functions. Positive definiteness in the estimate is therefore automatically enforced, and the technique can be implemented with standard regression packages. No kernel choice is needed in practical implementation, but basis functions need to be chosen and a smoothing parameter corresponding to the number of basis functions needs to be selected. An automated approach to making this selection based on optimizing the asymptotic mean squared error is derived. The limit theory of the new estimator shows that its properties, including the convergence rate, are comparable to those of conventional HAC estimates constructed from quadratic kernels.My thanks go to Bruce Hansen, Guido Kuersteiner, and two referees for comments on an earlier version of the paper. NSF research support under grant SES 00-92509 is acknowledged.

This paper is concerned with the practical problem of conducting inference in a vector time series setting when the data are unbalanced or incomplete. In this case, one can work with only the common sample, to which a standard HAC/bootstrap theory applies, but at the expense of throwing away data and perhaps losing efficiency. An alternative is to use some sort of imputation method, but this requires additional modeling assumptions, which we would rather avoid. We show how the sampling theory changes and how to modify the resampling algorithms to accommodate the problem of missing data. We also discuss efficiency and power. Unbalanced data of the type we consider are quite common in financial panel data; see, for example, Connor and Korajczyk (1993, Journal of Finance 48, 1263–1291). These data also occur in cross-country studies.I thank Greg Connor, Esfandiar Maasoumi, Peter Phillips, Peter Robinson, Yoon Whang, and two referees for helpful comments.

This paper proposes a new class of heteroskedastic and autocorrelation consistent (HAC) covariance matrix estimators. The standard HAC estimation method reweights estimators of the autocovariances. Here we initially smooth the data observations themselves using kernel function–based weights. The resultant HAC covariance matrix estimator is the normalized outer product of the smoothed random vectors and is therefore automatically positive semidefinite. A corresponding efficient GMM criterion may also be defined as a quadratic form in the smoothed moment indicators whose normalized minimand provides a test statistic for the overidentifying moment conditions.

Smoothed nonparametric estimates of the spectral density matrix at zero frequency have been widely used in econometric inference, because they can consistently estimate the covariance matrix of a partial sum of a possibly dependent vector process. When elements of the vector process exhibit long memory or antipersistence such estimates are inconsistent. We propose estimates which are still consistent in such circumstances, adapting automatically to memory parameters that can vary across the vector and be unknown.

This paper provides a survey on a class of so-called subspace methods whose main proponent is CCA proposed by Larimore (1983, Proceedings of the 1983 American Control Conference 2). Because they are based on regressions these methods for the estimation of ARMAX systems are attractive as a result of their conceptual simplicity and their numerical advantages as compared to traditional estimators based on criterion optimization. Under the assumption of correct specification the methods provide consistent and asymptotically normal estimates for stationary ARMAX processes where the innovations may be conditionally heteroskedastic and the exogenous inputs are strictly stationary of sufficiently short memory. For stationary autoregressive moving average (ARMA) processes with independent and identically distributed (i.i.d.) Gaussian innovations the estimates are even asymptotically efficient. For I(1) ARMA processes the estimates of both the long-run and the short-run dynamics are consistent without using the knowledge that the data are integrated in the algorithm. Additionally the algorithms provide easily accessible information on the appropriateness of the chosen model complexity. The algorithms include a number of design parameters that have to be set by the user. These include the order of the estimated system. This paper collects up-to-date knowledge on the effects of these design parameters, leading to a number of suggested automated choices to obtain a fully automated estimation procedure.

This paper considers the problems facing decision makers using econometric models in real time. It identifies the key stages involved and highlights the role of automated systems in reducing the effect of data snooping. It sets out many choices that researchers face in construction of automated systems and discusses some of the possible ways advanced in the literature for dealing with them. The role of feedbacks from the decision maker's actions to the data generating process is also discussed and highlighted through an example.

This paper uses the specific-to-general methodological approach that is widely used in science, in which problems with existing theories are resolved as the need arises, to illustrate a number of important developments in the modeling of univariate and multivariate financial volatility. Some of the difficulties in analyzing time-varying univariate and multivariate conditional volatility and stochastic volatility include the number of parameters to be estimated and the computational complexities associated with multivariate conditional volatility models and both univariate and multivariate stochastic volatility models. For these reasons, among others, automated inference in its present state requires modifications and extensions for modeling in empirical financial econometrics. As a contribution to the development of automated inference in modeling volatility, 20 important issues in the specification, estimation, and testing of conditional and stochastic volatility models are discussed. A “potential for automation rating” (PAR) index and recommendations regarding the possibilities for automated inference in modeling financial volatility are given in each case.

In Perez-Amaral, Gallo, and White (2003, Oxford Bulletin of Economics and Statstics 65, 821–838), the authors proposed an automatic predictive modeling tool called relevant transformation of the inputs network approach (RETINA). It is designed to embody flexibility (using nonlinear transformations of the predictors of interest), selective search within the range of possible models, control of collinearity, out-of-sample forecasting ability, and computational simplicity. In this paper we compare the characteristics of RETINA with PcGets, a well-known automatic modeling method proposed by David Hendry. We point out similarities, differences, and complementarities of the two methods. In an example using U.S. telecommunications demand data we find that RETINA can improve both in- and out-of-sample over the usual linear regression model and over some models suggested by PcGets. Thus, both methods are useful components of the modern applied econometrician's automated modeling tool chest.

This paper presents a set of questions prepared by Clive Granger with responses by David Hendry on the use of PcGets (see Hendry and Krolzig, 2001) in data modeling and as a new research tool. PcGets is an Ox package (see Doornik, 2001) implementing automatic general-to-specific (Gets) modeling for linear regression models based on the theory of reduction, as in Hendry (1995, Ch. 9).

Since the paper by David Hendry and me, which is at the center of this discussion, was written by us separately, I think it is worth recording how this paper started. David was attending the Allied Social Sciences Association meeting in San Diego in January 2004 and stayed with my wife and me. In preparing a talk I realized that over my career I had been fortunate enough to have about eighty coauthors, but I found that the list did not include David. As he had just discussed the use of PcGets as a research tool, I proposed a paper exploring the limits of such a tool. The paper was soon born, and, after the almost obligatory journal rejection, it was picked up by Peter Phillips and this special issue resulted.