‘Just as early chess-playing programs were easily defeated, but later ones could beat
grandmasters, so we anticipate computer-automated model selection software developing
well beyond the capabilities of the most expert modellers. Deep Blue may be just round the
corner.’
That is the conclusion of Professors David Hendry and Hans-Martin Krolzig, writing in the
Economic Journal.
There are many examples where such tools are already used, including at central banks for
modelling macroeconomic relationships and investigating lagged reactions. Other potential
settings include those with many possible variables, as in actuarial studies, or interactions
between variables, as in advertising markets.
Economics advances by an interaction between theoretical ideas and empirical evidence.
But theories are usually simplified, incomplete and evolve over time; whereas data are less
than plentiful, often imperfectly measured and reflect an ever-changing economy.
Consequently, many aspects of the formulation of empirical models are uncertain,
necessitating some form of model selection procedure.
Unfortunately, the process of selecting a model from empirical evidence has itself been
subject to considerable doubt, often deemed little better than mindless ‘data mining'.
Moreover, theoretical analyses of the properties of complicated selection procedures have
proved intractable. Despite such problems, over the last few years, major advances have
occurred in the theory and practice of automatic model selection methods.
The approach discussed by Hendry and Krolzig is embodied in software called PcGets,
where Gets denotes general-to-specific modelling. This commences from the most general
model thought necessary to characterise the data evidence consistent with the theoretical
framework, then simplifies it to a parsimonious form that still captures all the relevant
information.
By implementing automatic selection in the Gets methodology, itself of long standing in
economics, computer simulation studies of that approach also become feasible. These
have confirmed that the operational characteristics of such selection algorithms are in fact
excellent across a wide range of states of nature.
In particular, when there is a correct specification, that model is selected almost as often in
practice as the maximum suggested by probability derivations. More generally, ‘over-fitting’
is not a major difficulty. Surprisingly, nearly unbiased parameter estimates can also be
obtained, with relatively accurate estimates of their uncertainty.
The basis for these advances is explained, and the authors describe further important
developments, including handling what have been thought until recently to be intractable
problems. Examples of these are: where there are more variables than observations in
regressions; where such variables are perfectly collinear; modelling simultaneous equations
without prior identifying restrictions; selecting cointegration relations in integrated
processes; checking for general forms of parameter non-constancy; and determining
specifications for non-linear models.
ENDS
Notes for Editors: ‘The Properties of Automatic GETS Modelling’ by David Hendry and
Hans-Martin Krolzig is published in the March 2005 Economic Journal.
Hendry is Professor of Economics at Oxford University; Krolzig is Professor of Economics
at the University of Kent.
For Further Information: contact David Hendry on 01865-278554 (email: via
maureen.baker@nuffield.ox.ac.uk); or RES Media Consultant Romesh Vaitilingam on 0117-
983-9770 or 07768-661095 (email: romesh@compuserve.com).