15
interesting insights on the out-of-sample performance of the candidate forecast
models (Table 8). For the first German sample, the FB test statistics exhibit rela-
tive large negative values, i.e. the null hypothesis that the out-of-sample errors are
not worse than those in sample is not rejected. This is particularly true for the
sMed - and the sDiff -indicator. For the second sub-period, however, FB tests indi-
cate forecast breakdowns for the sMed -, the sDiff - and the str20 -indicator-based
models. This evidence is consistent with our finding that one might fruitfully em-
ploy sentiment indicators for the purpose of forecasting inflation, emphasizing that
their contribution will tend to be smaller in stable times.
[Table 8: about here]
5. Conclusions
In this paper we construct several indicators capturing whether a given inflation
rate is the result of price increases for many components in the CPI basket or
rather the consequence of price hikes for a relatively small number of goods and
services with a high weight in the basket. Since inflation is supposed to be per-
ceived more intensively in the first case, we label our indicators inflation senti-
ment. We also demonstrate that simple sentiment indicators are highly correlated
with differences of familiar core measures and headline inflation.
With regard to the US we find that inflation sentiment indicators tend to improve
the accuracy of inflation forecasts as measured by the RMSFE by roughly 20%,
compared to a standard Phillips curve approach. The differences are significant
according to the Diebold-Mariano and modified Diebold-Mariano tests for fore-
cast horizons up to eight quarters. Here, indicators based on familiar core meas-
ures show forecast accuracy similar to our new indicators. Furthermore, a forecast
breakdown test indicates that the out-of-sample forecast errors of the alternative
candidate models do not deviate significantly from their in-sample fit, suggesting
that the forecasts based on our indicators are more stable than the standard Phillips
curve approach.
The results derived on the basis of German data are less uniform. Their heteroge-
neity is indicative for the role of circumstances: inflation sentiment indicator-
based forecasts seem to be particularly powerful, if inflation is volatile as it has
been the case in the 1985:1 to 1998:4 sample of West German data. The RMSFE
is reduced up to an half. In the 1993:1 to 2007:4 period, tough, when inflation was
quite stationary, the sentiment indicators do not outperform Phillips curve-based
forecasts. However, also under these circumstances encompassing tests suggest
that our inflation sentiment indicators add valuable information. For this German
data the performance of core inflation-based indicators appears to be worse than
that of the new indicators. Since the latter require less data - neither explicit