
 Start

4010
 Prefix

Given that the estimated output gap is used as an indicator to measure the extent of
inflationary pressures in the economy, we evaluate the use of such models for the implementation
of monetary policy. Our starting point is the
 Exact

Gerlach and Smets (1997)
 Suffix

unobservedcomponents
model, which they applied to the G7 countries. After subjecting this model to various diagnostic
tests, we modify certain assumptions in it to reflect specific aspects of the Canadian economy.
 (check this in PDF content)

 Start

5329
 Prefix

indicateur de l’intensité des pressions inflationnistes
au sein de l’économie, l’auteure de l’étude cherche à évaluer si le recours aux modèles espace
d’états pour estimer la production potentielle et, partant, l’écart de production peut être d’une
quelconque utilité dans la mise en oeuvre de la politique monétaire. Le point de départ de son
analyse est le modèle à composantes non observées que
 Exact

Gerlach et Smets (1997)
 Suffix

appliquent aux
pays du Groupe des Sept. Après avoir soumis le modèle à plusieurs tests de diagnostic, l’auteure
modifie certaines des hypothèses initiales pour tenir compte d’aspects précis de l’économie
canadienne.
 (check this in PDF content)

 Start

8337
 Prefix

The extended multivariate filter (EMVF)
used at the Bank of Canada in the quarterly projection model is such a model.1Finally, the authors
examine filtering models based on vector autoregressions (VARs). Among these are the structural
VAR methods of
 Exact

Blanchard and Quah (1989),
 Suffix

the multivariate BeveridgeNelson decomposition
and Cochrane’s (1994) methodology.2
StAmant and van Norden show that both univariate mechanical filters, as well as their
hybrid extensions, may have problems in adequately separating output into trend and meanreverting components.3 These filters perform poorly in extracting business cycle frequencies of 6
to 32 quarters from series such as real GDP
 (check this in PDF content)

 Start

8875
 Prefix

both univariate mechanical filters, as well as their
hybrid extensions, may have problems in adequately separating output into trend and meanreverting components.3 These filters perform poorly in extracting business cycle frequencies of 6
to 32 quarters from series such as real GDP which have an important permanent component. In
addition, they are subject to severe endofsample problems.
1.See
 Exact

Butler (1996).
 Suffix

2.An example of the application of these methods to U.S. data can be found in Dupasquier, Guay and StAmant
(1999).
3.See also Harvey and Jaeger (1993) and King and Rebelo (1993).
As one alternative to the above class of models, StAmant and van Norden suggest the use
of structural VAR (SVAR) methods for measuring trend output.4 These models are favoured
because they are not confronted with endof
 (check this in PDF content)

 Start

8967
 Prefix

adequately separating output into trend and meanreverting components.3 These filters perform poorly in extracting business cycle frequencies of 6
to 32 quarters from series such as real GDP which have an important permanent component. In
addition, they are subject to severe endofsample problems.
1.See Butler (1996). 2.An example of the application of these methods to U.S. data can be found in
 Exact

Dupasquier, Guay and StAmant (1999).
 Suffix

3.See also Harvey and Jaeger (1993) and King and Rebelo (1993).
As one alternative to the above class of models, StAmant and van Norden suggest the use
of structural VAR (SVAR) methods for measuring trend output.4 These models are favoured
because they are not confronted with endofsample issues.
 (check this in PDF content)

 Start

9043
 Prefix

In
addition, they are subject to severe endofsample problems.
1.See Butler (1996). 2.An example of the application of these methods to U.S. data can be found in Dupasquier, Guay and StAmant (1999). 3.See also Harvey and Jaeger (1993) and
 Exact

King and Rebelo (1993).
 Suffix

As one alternative to the above class of models, StAmant and van Norden suggest the use
of structural VAR (SVAR) methods for measuring trend output.4 These models are favoured
because they are not confronted with endofsample issues.
 (check this in PDF content)

 Start

9732
 Prefix

However, the authors also caution that confidence intervals around the estimated gap are generally
wide and that results are sensitive to the variables included in the estimated system.
Other reasons for being cautious about the interpretation of the implied dynamics obtained
from SVARs are given in
 Exact

Cooley and Dwyer (1998).
 Suffix

These authors explain that, since the
identification of SVARs is conditional on a set of atheoretical assumptions which cannot be tested,
results from the SVAR are quite sensitive to misspecification in these assumptions.
 (check this in PDF content)

 Start

11234
 Prefix

the equations in the statespace system on the other, we can still analyze a welldefined macroeconomic system and
interpret the parameters and residuals in a straightforward fashion.6(ii) The capacity of the model
to provide us directly with confidence intervals around the measured gap or potential.7 (iii) The
4.Another is the TOFU method proposed in van Norden (1995).
5.For more details, see
 Exact

Phillips (1989).
 Suffix

6.Ideally, it would be desirable to make use of a fully specified structural and stochastic model. Unfortunately,
this also is problematic as it necessitates, among other things, making assumptions on the equilibrium level of
the various components used in the production function, as well as its aggregated form.
 (check this in PDF content)

 Start

11907
 Prefix

Studies employing
these types of models frequently encounter difficulties in finding suitable data for their longrun values.
7.This is the mean square error associated with the estimated state vector and is obtained as a byproduct of
applying the Kalman filter. It is also known as the filter uncertainty of the model. However, as pointed out by
 Exact

Hamilton (1986),
 Suffix

two types of uncertainties are associated with these models: filter uncertainty and parameter
uncertainty. If one adopts the Bayesian perspective that the true value of the state vector is random, then our
knowledge of it, based on observable variables, is reflected in a probability distribution.
 (check this in PDF content)

 Start

13079
 Prefix

uncertainty, must be added to the above.
fact that we can also directly obtain outofsample forecasts on the observable variables in the
model, which provide an easy way of assessing the goodness of fit of the adopted specification
and of ensuring that our estimates are useful in the formulation of policy.
The statespace framework was first used to estimate trend output on U.S. data by
 Exact

Kuttner (1994).
 Suffix

It was subsequently extended by Gerlach and Smets (1997) and applied to the G7
countries. For our part, we examine the implications of using different versions and extensions of
the Gerlach and Smets (1997) model (hereafter GS) for Canada.
 (check this in PDF content)

 Start

13126
 Prefix

fact that we can also directly obtain outofsample forecasts on the observable variables in the
model, which provide an easy way of assessing the goodness of fit of the adopted specification
and of ensuring that our estimates are useful in the formulation of policy.
The statespace framework was first used to estimate trend output on U.S. data by Kuttner (1994). It was subsequently extended by
 Exact

Gerlach and Smets (1997) and
 Suffix

applied to the G7
countries. For our part, we examine the implications of using different versions and extensions of
the Gerlach and Smets (1997) model (hereafter GS) for Canada. In all cases, the output gap is
defined as the component of output which is consistent with change in the rate of inflation, all
other things being equal.
 (check this in PDF content)

 Start

13274
 Prefix

The statespace framework was first used to estimate trend output on U.S. data by Kuttner (1994). It was subsequently extended by Gerlach and Smets (1997) and applied to the G7
countries. For our part, we examine the implications of using different versions and extensions of
the
 Exact

Gerlach and Smets (1997)
 Suffix

model (hereafter GS) for Canada. In all cases, the output gap is
defined as the component of output which is consistent with change in the rate of inflation, all
other things being equal.
The organization of the paper is as follows.
 (check this in PDF content)

 Start

18859
 Prefix

Notice that this is a very general statespace framework so that when no ARCH errors are
present in the system, the above filter reduces to the wellknown Kalman filter.
3.The Gerlach and Smets model
We now move to the presentation of the
 Exact

Gerlach and Smets (1997)
 Suffix

model which is a special case
of the general model presented above. As mentioned, it is a modification of the Kuttner (1994)
model which was originally applied to the United States and where potential output and the slope
of the Phillips curve were simultaneously estimated in an unobserved components framework.
 (check this in PDF content)

 Start

18993
 Prefix

Notice that this is a very general statespace framework so that when no ARCH errors are
present in the system, the above filter reduces to the wellknown Kalman filter.
3.The Gerlach and Smets model
We now move to the presentation of the Gerlach and Smets (1997) model which is a special case
of the general model presented above. As mentioned, it is a modification of the
 Exact

Kuttner (1994)
 Suffix

model which was originally applied to the United States and where potential output and the slope
of the Phillips curve were simultaneously estimated in an unobserved components framework.
The log of quarterly real output,, is assumed to be the sum of log real potential output,
, and a log cyclical component,.
 (check this in PDF content)

 Start

20266
 Prefix

These variables are included to capture the effects of temporary relative price shocks on inflation.
Finally, the error terms are assumed to be mean zero and normally distributed. They enter the
11.This is the assumption made by
 Exact

Watson (1986).
 Suffix

It has since been used in a standard fashion in various
macroeconomic models.
inflation equation as a moving average process to capture any remaining inertia in supply shock
variables. This also allows the model to be identified.
 (check this in PDF content)

 Start

27727
 Prefix

However, it remains an important question and
is left for future research.
15.In fact, if the true process is I(1), but instead an I(2) is assumed, the model may not be wellidentified. This
adversely affects any inference results from the model.
16.See, for instance,
 Exact

Duguay (1994).
 Suffix

17.One such study is by Laxton, Rose and Tetlow (1993).
18.An example is the study by Ricketts and Rose (1995) for Canadian inflation data. See also Fillion and Léonard
(1997) for an estimated Phillips curve with various regimes.
4.
 (check this in PDF content)

 Start

27765
 Prefix

However, it remains an important question and
is left for future research.
15.In fact, if the true process is I(1), but instead an I(2) is assumed, the model may not be wellidentified. This
adversely affects any inference results from the model.
16.See, for instance, Duguay (1994). 17.One such study is by
 Exact

Laxton, Rose and Tetlow (1993).
 Suffix

18.An example is the study by Ricketts and Rose (1995) for Canadian inflation data. See also Fillion and Léonard
(1997) for an estimated Phillips curve with various regimes.
4.Extensions and model applicability to Canada
4.1Fine tuning
In Table 1 parameters significant at the 10percent level are in bold.
 (check this in PDF content)

 Start

27826
 Prefix

However, it remains an important question and
is left for future research.
15.In fact, if the true process is I(1), but instead an I(2) is assumed, the model may not be wellidentified. This
adversely affects any inference results from the model.
16.See, for instance, Duguay (1994). 17.One such study is by Laxton, Rose and Tetlow (1993). 18.An example is the study by
 Exact

Ricketts and Rose (1995)
 Suffix

for Canadian inflation data. See also Fillion and Léonard
(1997) for an estimated Phillips curve with various regimes.
4.Extensions and model applicability to Canada
4.1Fine tuning
In Table 1 parameters significant at the 10percent level are in bold.
 (check this in PDF content)

 Start

27889
 Prefix

This
adversely affects any inference results from the model.
16.See, for instance, Duguay (1994). 17.One such study is by Laxton, Rose and Tetlow (1993). 18.An example is the study by Ricketts and Rose (1995) for Canadian inflation data. See also
 Exact

Fillion and Léonard (1997)
 Suffix

for an estimated Phillips curve with various regimes.
4.Extensions and model applicability to Canada
4.1Fine tuning
In Table 1 parameters significant at the 10percent level are in bold. Looking at Column 4 we can
see that many of the coefficients are not significant.
 (check this in PDF content)

 Start

33387
 Prefix

This supposes that the observed
20.Since these models are nonnested, we do not use likelihood ratio tests to assess the overall goodness of fit.
Instead, we rely on indirect criteria to select the best specification.
21.For more detail, see
 Exact

Stuber (1986).
 Suffix

slowdown took place very gradually. Yet, an explanation based on a one or twotime break in the
drift term is equally likely.
Opinions diverge as to the right explanation, but, at this stage, both appear equally valid.
 (check this in PDF content)

 Start

34115
 Prefix

This debate has
been ongoing for a long time now and has spawned numerous studies on methods of testing for
break points when the break time is difficult to pin down.
The latest in this category is the extensive work by
 Exact

Bai and Perron (1998).
 Suffix

These authors
discuss the properties of least squares estimators in a wide class of linear regression models in the
presence of multiple structural breaks with unknown break points. These models include cases
where the residuals are autocorrelated and heteroskedastic and where there are lagged dependent
variables in the regressors.
 (check this in PDF content)

 Start

38805
 Prefix

the purpose of the output gap measure is consistency with short
term inflation forecasting, and taking into account the relative complexity of our framework, we
will not be addressing this issue in the present study. Of more importance is the observed
autocorrelation in the inflation equation residuals in all of the models. These effects could be due
to misspecified expectations. For instance,
 Exact

Ricketts and Rose (1995),
 Suffix

using a Markovswitching
model, found three distinct regimes for Canadian inflation between 1961 and 1993. In their model
both means and variances change with the state of the economy. If agents know which inflation
regime they are in when they make their forecasts, inflation expectations will be different in each
state.
 (check this in PDF content)

 Start

40662
 Prefix

In
order to account for this fact, we assign different dummy variables to these two time periods. This
amounts to specifying a total of four distinct regimes for inflation, and consequently, of inflation
expectations.
22.The construction of the dummy variables is similar to the method followed by
 Exact

Fillion and Léonard (1997).
 Suffix

Second, we reestimate the statespace model with a unit root drift and core inflation
having imposed these four distinct regimes for inflation.23 In all regimes, and to preserve
parsimony, core inflation is assumed to follow a stationary autoregressive process of the second
order.24
Estimation results for the models with nominal and real exchange rates are tabulated in
Table 5.
 (check this in PDF content)

 Start

50657
 Prefix

This model has a good overall fit, plausible parameter estimates, and the best
forecast performance.
5.Conclusion
In this paper, we assess the implications of using statespace models for measuring potential
output (and the output gap) in Canada. The models examined are different versions or extensions
of the unobservedcomponents model of
 Exact

Gerlach and Smets (1997).
 Suffix

These modifications were
motivated partly in response to existing empirical findings on Canadian data, and partly because
of the results of diagnostics checking on the basic model.
First we focus on the assumed specification for the output equation of the model and apply
the Bai and Perron (1998) structural change test with unknown change point.
 (check this in PDF content)

 Start

50955
 Prefix

These modifications were
motivated partly in response to existing empirical findings on Canadian data, and partly because
of the results of diagnostics checking on the basic model.
First we focus on the assumed specification for the output equation of the model and apply
the
 Exact

Bai and Perron (1998)
 Suffix

structural change test with unknown change point. Since we find that
we cannot reject a onetime break in the drift of output growth, we estimate versions of the model
with a breaking drift. We also estimate a version of the GS model with a modified error structure
after detecting the presence of ARCH effects in this equation.
 (check this in PDF content)