12
4. Results
4.1 United States
A relatively long time series is available for the US. During the sampling period,
the US economy was relatively stable in terms of its underlying structure. The root
mean squared forecast errors (RMSFE) of the standard Phillips curves serve as the
relevant benchmark throughout the analysis9. In the estimates the real side of the
US economy is either represented by changes in GDP or in the unemployment
rate. Table 3 summarizes the RMSFE for the Phillips curve and compares it to the
results of the five alternative specifications featuring inflation sentiment indicators
instead of inflation. We can be confident that the results do not depend on the
choice of the real side variable: Differences of the unemployment rate and changes
in GDP generate more or less the same results concerning the relative RMSFE. In
both cases, the sMed - and the s Diff -indicators improve forecast accuracy for all
forecast horizons. For short horizons the improvement achieved is about 10% on
average. For longer forecast horizons the improvement climbs up to 30%. The
s Diff -indicator even performs slightly better than the more established sMed -
measure. All differences to the benchmark forecasts are statistically significant,
some of them at a 99% level. However, by replacing the inflation differences by
the sMom -indicator, we get statistically worse results.
Since our inflation sentiment indicators are highly correlated with the differences
between the two core inflation measures and headline inflation, the forecast poten-
tial of these core measures should be very similar to the proposed inflation senti-
ment indicators. From Table 3 it becomes obvious that the weighted median-
indicator, sWMed , performs worse than its un-weighted pendant up to a forecast
horizon of roughly one year and better if longer forecast horizons are considered.
The trimmed mean-indicator, str20 , displays high forecast accuracy in particular
for longer forecast horizons, one the one hand. On the other hand, the differences
in the RMSFE are only marginal and the DM and modified DM test statistics
show lower absolute values than in the inflation sentiment indicator models. All in
all, both concepts seem to have a similar forecasting power for US inflation.
[Table 3: about here]
Since forecast models using sMed , s Diff , sWMed and str20 perform better than that
using the inflation rate itself, the results of the encompassing tests are not surpris-
ing. Table 4 confirms that the null hypothesis that the standard forecasting model
9 Additionally we tested whether the Phillips curve also can outperform a naïve forecast. For
the U.S we found the Phillips curve showing smaller forecast errors than e.g. univariate fore-
casts or using the sample mean as predictor. In most cases, the DM and the MDM test show
that the differences are significant. The results are available from the authors upon request.