increase in the price of crude oil.
However, not all commodity prices
have the same impact on inflation,
since certain goods represent only a
small portion of world consumption
and production. In addition, equal
weighting tends to overstate the im-
portance of groups of commodities.
For example, as figure 1 shows, the
CRB index is heavily weighted toward
agricultural commodities, whose fu-
tures prices are constantly affected by
changing weather reports. As a result,
the CRB responds sharply to price
swings in commodities such as coffee
and cotton that have very little impact
on overall inflation.
An alternative approach would be to
weight each component in propor-
tion to its relative value in world pro-
duction. The Producer Price Index,
for example, uses a production-based
weighting scheme for its components,
where weights depend on the product
output value at the time of shipment
to another industry. The higher the
output value of the commodity, the
heavier its weight in the index. Simi-
larly, under a world production
weighting scheme, crude oil, for ex-
ample, would have three times the
weight it now has in CRB, while cocoa
would have 1/24 the weight used in
CRB. This method of weighting
would reflect the fact that a sustained
increase in the price of crude oil has a
larger impact on overall inflation than
a comparable increase in the price of
cocoa. This is because crude oil is an
input to a vastly larger number of
finished goods and has a much great-
er world production value than cocoa.
JOCCI uses yet another weighting
scheme that gives more importance to
materials whose price movements are
believed to lead consumer price infla-
tion. This is Consistentwith the basic
idea that among commodities used
intensively in cyclical industries, pric-
es tend to increase before consumer
prices do. Theoretically, this weight-
ing scheme should eliminate some of
the problems of equal weighting and
increase the indicator’s ability to an-
ticipate inflation.
Compositional issues such as these
make commodity price indexes sus-
ceptible to sharp fluctuations, since
materials prices respond not only to
economic fundamentals but also to
various market forces. Pindyck and
Rotemberg, for example, found that
prices of unrelated commodities tend
to move together as a result of “herd”
behavior in financial markets.3 That
is, traders seem to exhibit a similar
behavior in all commodities markets
instead of responding to specific eco-
nomic events. Thus, for instance,
futures prices of precious metals have
been responding to movements in
grain futures, which are affected by
constantly changing weather fore-
casts. Clearly, prices of precious met-
als should not be affected by weather
conditions. But when grain prices
rise, CRB also increases because it is
heavily weighted toward agricultural
commodities. Traders in other com-
modities markets fear higher inflation
and react accordingly. Such behavior
is reasonable if the index’s increase is
truly signaling higher inflation. It is
not reasonable, however, if move-
ments in the index are caused by
relative price changes. Given the
many compositional quirks of the
various commodity price indexes, it is
very difficult to determine whether an
increase in an index is supply-driven
or actually indicates inflation.
How well do they forecast inflation?
Do commodity price indexes help
forecast inflation? That is, if such an
index were included in a forecasting
model containing data on past infla-
tion, would the resulting forecast be
more accurate than the
one the model would
have generated without
the index? We attempt-
ed to answer this ques-
tion by comparing his-
torical data on actual
inflation with the fore-
casts the commodity
price indexes would
have generated for the
same periods.
We evaluated the com-
modity price indexes in
three steps. First, we
produced inflation fore-
casts from January 1970
to June 1994 based only
on past inflation and calculated the
average size of the forecast errors over
this period, as measured by root mean
squared errors (RMSEs).4 In this first
step, we used a simple autoregressive
model which we called the no-indica-
tor model, with 12 lags of inflation
growth and a constraint term on the
right-hand side of the equation. Next,
we repeated this analysis by adding
one commodity price index to the no-
indicator model to produce bivariate
models which we called indicator
models. We tested three such models,
each including one of the three com-
modity price indexes; in all of these
models, both the index and inflation
growth were lagged 12 months.
Third, we compared the average fore-
cast error from each indicator model
with the average forecast error from
the no-indicator model. If the aver-
age error from an indicator model
was significantly smaller than the
average error from the no-indicator
model, then we would say that the
added index improved the forecast.
To quantify the statistical significance
of any apparent improvement in fore-
cast, we performed a t-test on the
difference between the two models’
squared forecast errors.
Figure 2 ranks the indicators accord-
ing to their average forecast errors at
3-month, 6-month, and 12-month
forecast horizons. In simple terms,
the lower the average forecast error,
the better the performance of the
forecasting model. JOCCI and SMPS
seemed to perform better than the
no-indicator and CRB models at all
2. Average forecast errors
3-month 6-month 12-month
Indicator |
RMSE Rank |
RMSE Rank |
RMSE Rank | |||
None |
2.269 |
4 |
2.098 |
3 |
2.214 |
3 |
CRB |
2.264 |
3 |
2.112 |
4 |
2.214 |
4 |
JOCCI |
2.171 |
1 |
1.950 |
1 |
2.085 |
2 |
SMPS |
2.203 |
2 |
1.959 |
2 |
2.038 |
1 |
Significance levels | ||||||
CRB |
0.934 |
0.815 |
0.993 | |||
JOCCI |
0.165 |
0.069 |
0.110 | |||
SMPS |
0.280 |
0.057 |
0.026 |
Note: RMSEs are root mean squared errors. Significance levels
were for the t-test of the null hypothesis that the mean of the
difference of the squared errors was equal to zero.