David F. Hendry 17(2000, 2001) and Johansen (2006). Structural breaks matter most when they induce
location shifts in the processes under analysis, but those can also be removed in
part by differencing or co-breaking (see Hendry and Massmann, 2007). Forecast
failure – defined as a significant deterioration in forecast performance relative to its
anticipated outcome, usually based on historical performance – is all too common,
and is almost certainly due to structural breaks (see, e.g., Clements and Hendry,
1998, 1999, 2002b).
Because much observed data variability is due to factors absent from economic
theories, a serious gap exists between macroeconomic theory models and applied
econometric findings (see Spanos, 1989; Juselius, 1993; Hendry, 1995b; Nymoen,
2002). All economic theories rely on implicitceteris paribusclauses, as “controls
in thought experiments,” although in a general equilibrium system in which
everything depends on everything else,ceteris paribusis suspect. In empirical
modeling,ceteris paribuscannot apply under non-stationarity even if the rele-
vant variables are strongly exogenous, since “other things” will not be “equal.”
Cartwright (2002) describesceteris paribusas roughly equivalent to “if nothing
interferes then...some regularity is observed.” In non-stationary processes, noth-
ing will interfere only if all other factors are irrelevant, not because they will not
change. Many sources of wide-sense non-stationarity impinge on economic data,
including technical progress, R&D, new legislation, institutional changes, regime
shifts, financial innovation, shifting demography, evolving social and political
mores, as well as conflicts and other major catastrophes, inducing both evolu-
tion and structural breaks, all of which change the distributional properties of
data.
Two resolutions are possible to wide-sense non-stationarity. First, a “minor influ-
ence” theorem could show on theoretical or evidential grounds that all omitted
factors can be neglected, either because changes in them are of a smaller order
of importance than included effects, or because they are orthogonal to all the
effects that matter (see Hendry, 2005, and compare Boumans, 2005, who refers
toceteris neglectisandceteris absentibus). Neither condition is plausible unless at
least all the major influences have been included. Doing so brings us anyway
to the second solution, namely including all potentially relevant variables at the
outset, embedding theory models in more general systems that also allow for all
the empirically-known influences, as well as the many historical contingencies
that have occurred. Thus, institutional knowledge and economic history become
essential ingredients in Applied Econometrics. Far from diminishing the impor-
tance of economic reasoning as a basis for empirical econometrics, including all
non-stationarities seems the only way to reveal the underlying economic behav-
ior uncontaminated by excluded changes. Of course, theory models of the likely
behavioral reactions of economic agents to major changes would also help. As
macro-data are the aggregates of the economic microcosm, these problems must
afflict all empirical econometric studies, and are not merely a problem for time
series analysts. Since the need to model all non-stationarities if empirical results
are to be useful is important for both economics and econometrics, the next section
considers its prevalence.