The Economist November 6th 2021 GraphicdetailInflation 81
flaws. Changes in food and energy prices
are not necessarily unusually large or
shortlived. And trimmed means’ weight
ing schemes are plagued by abrupt cliffs. In
the Cleveland Fed’s version, which lops off
the top and bottom 8% of the index, an
item in the 93rd percentile when sorted by
price changes is removed entirely, whereas
one in the 92nd gets its full weight.
With this in mind, we have devised an
alternative inflation index. Like trimmed
means, it adjusts items’ weights based on
their recent price changes. But its weights
are shaped like a smooth hill rather than a
box. Components with inflation near the
median get the most emphasis, and those
with the biggest price changes get the least.
Our hill looks a bit like Uluru in Austra
lia: a broad central plateau, flanked by a
steep slope on the left side and a gentler
one on the right. (The Dallas Fed’s trimmed
mean is also asymmetric, counteracting
bias caused by pricechange distributions’
lopsidedness.) Most items with negative or
low inflation get a hefty weight; those
whose prices are rising fastest count for
25% as much as those in the middle do.
When using the past year of data to
predict pceinflation during the following
year, this method is more accurate than ei
ther using core inflation or expecting infla
tion to remain constant. Since 1959, its
oneyear forecasts have also outperformed
those of the Dallas Fed’s trimmed mean.
Some of this apparent advantage stems
from the design of our study: the Dallas Fed
sought to maximise accuracy for different
time periods and forecast horizons than
ours. However, its trimmed mean’s errors
in the 1970s illustrate the risks of such a
deep trim. Amid two oil shocks, some of
the fastestrising prices—those of energy
related items—just kept rising. The less
weight an index placed on such goods, the
worse it predicted inflation one year out.
The current episode of concentrated in
flation differs from the 1970s in many
ways. Surging prices for goods like home
appliances are unlikely to feed through to
other costs, as oil prices do. And in general,
treating outliers the same as more repre
sentative items has been a mistake. But in
cases where such price changes foreshad
ow broader supply constraints, ignoring
them entirely can be an even bigger error.
As a sensecheck of our “Uluru” meth
od, we also built a model that forecasts pce
inflation using only excess concentration
and the one and tenyear trailing inflation
rates. Both this approach and the Uluru in
dex yield a 4.1% prediction for inflation
during the next 12 months. That is below
the current level of 4.4%, but above the
Fed’s target of 2% and the 2.3% value of the
Dallas Fed’s trimmed mean for thepast 12
months. If this forecast comes true,inter
estrate rises will almost surely follow.n
→ Four ways to weight an inflation indexHeadline inflation*
WeightsitemsattheirshareofconsumerspendingCore inflation
ExcludesfoodandenergypricesTrimmed-meaninflation
Excludesproductswithbigpriceswings“Uluru”inflation
Smoothlyreweightsitems*Weightedaverageofeachitem’syear-on-yearchangeinpersonalconsumptionexpendituresindex
†Differenceinroot-mean-squareerrorpredictingheadlinePCEinflationduringthesubsequent 12 months→“Uluru”inflationislessvolatilethanheadline,anda betterpredictorHeadlinev “Uluru”inflationindices,%changeona yearearlierHousing
FoodPrescriptionsPrescriptionsPetrolFoodPetrolFood PetrolPetrolHousingHousingHousingFood000.5xEqual to
headlineEqual to
headline1.5x0.5x1.5x2.0x2.5x← Lower Percentile of price change Higher →← Lower Percentile of price change Higher →← Lower Percentile of price change Higher →← Lower Percentile of price change Higher →Weight ratio, May 2021Weight ratio, May 2021↓Headlineclosertonextyear’sinflation†↑“Uluru”closertonextyear’sinflation†Headline index“Uluru” index-2.502.55.07.510.012.51960s 70s 80s 90s 2000s 10s 20sCovid-1
pandemicGlobal financial
crisisOil
shockOil
shockFed-induced
recessionSources: Bureau of Economic
Analysis; The Economist