Gunnar Bårdsen and Ragnar Nymoen 903small for forecast horizons 1, 2, 3 and 4: the MFEs are less than 0.25 percentage
points. Norges Bank’s inflation forecasts from theInflation Report/Monetary Policy
Reportsare more biased than AIF for horizons 2–8 quarters ahead. The biases of AIF
become markedly bigger for forecasts of length 7–10 quarters, and are not much
different from the bias of forecasts produced by Norges Bank. The second panel of
the figure shows the MSFEs, to which large forecast errors contribute more than
small errors. This measure gives more or less the same picture at the MFEs.
17.4.5.3 Ex postforecast evaluation and robustification
The discussion of Figure 17.9 illustrated the general point that, although EqCMs
forecast well when a process is difference-stationary, they are non-robust if there
are non-stationarities due to location shifts in the forecast period. In this particular
case, the mean of the rate of inflation seems to have changed, in 2004 or earlier,
and the DDD is a more adaptable forecasting mechanism than the EqCMs in the
case of the before-forecast structural break. The main benefit of DDDs for model-
based forecasting, where one wants to retain the causal information of the model,
is that DDDs can help the forecaster to robustify the EqCM forecasts, by intercept
corrections. The cost associated with DDDs is, of course, that the forecasts are
more “noisy” than EqCM forecasts, hence the forecast-error variances associated
with robust forecasts can become large, such as when the first difference of an
autoregressive process doubles the one-step forecast variance. In a model with one
or two endogenous variables this cost may not be much of an issue, but in a multi-
equation forecasting setting there may be a problem of extracting “signal from
noise” in practice. In the rest of this section we illustrate these issues by considering
system forecasts.
Figure 17.11 shows dynamic NAM forecasts for the period 2003(1)–2007(3).^18
The sample period for the estimation ends in 2002(4). Unlike the inflation fore-
casts in Figure 17.9 which are real-timeex anteforecasts, we now considerex post
forecasts, which are conditioned by the true values of the non-modeled variables.
Hence the forecasts are not influenced by location shifts in the non-modeled vari-
ables (foreign CPI inflation, e.g.), but they are subject to location shifts that are
due to changes in the means of the estimated cointegration relationships, or the
autonomous growth rates (captured by intercepts after subtraction of equilibrium
correction means).
Figure 17.11 serves as the reference case for discussion of the degree of adapta-
tion of NAM forecasts, and for the properties of more robust forecasting devices
that we investigate for comparison. There are no less than five possible instances
of location shifts in the forecast of these nine variables. First, there is sys-
tematic overprediction of the rate of inflation (we use CPI inflation here) over
the length of the forecast horizon. In the light of the 70% prediction interval,
inflation overprediction is significant (forecast failure) in 2003. Second, unem-
ployment is overpredicted and GDP growth in underpredicted for the three
quarters of 2007. Third, the short-term interest rate is significantly overpredicted in
2003(2)–2004(4). Fourth, real credit growth is underpredicted (significantly) from