Trimmed Mean CPI Inflation
In a recent speech Governor Kohn of the Federal Reserve System noted that
"Certainly, the recent data on consumer prices have been encouragingly consistent with the downward tilt to inflation that the FOMC has been expecting. However, we need to be cautious about extrapolating trends from a couple of months of data. The data themselves are noisy - subject to month-to-month variations that are unrelated to more-persistent developments. And we need to recognize that some of the very recent disinflation may represent one-time influences."
Separating transitory "noise" from the price data to reveal the more persistent inflation trend that the central bank hopes to control is difficult - and controversial. It's difficult because it is often unclear when a price change is transitory and when it signals a change in the inflation trend. It's controversial because ignoring certain price signals may be interpreted as a selective and perhaps biased interpretation of the data.
The most common approach to reducing the "noise" in the price data is the so-called Core CPI measure, which excludes the prices of food and energy items. This measure certainly eliminates two of the most volatile, or "noisy" components in the U.S. retail price data. But this approach does not address transitory price fluctuations in other components of the retail market basket. And because it systematically eliminates two components from the retail market basket, it may also bias the inflation measure - if there are long-term movements in the prices of food and energy relative to other goods and services.
An alternative approach to measuring underlying or "core" inflation uses trimmed-mean estimators, which eliminate any price change above or below a certain threshold, regardless of what the component is. In the CPI, for example, one commonly sees a significant portion of the items in the market basket with price increases well above or below what the inflation trend ultimately is revealed to be.
Where the appropriate threshold lies is unclear, as is the question of which items get excluded from consideration. We can exclude the most extreme, say 5 percent, from each tail of the price-change distribution, resulting in a "10 percent trimmed-mean" inflation estimate. Or we could calculate a more substantial trim - say 49-1/2 percent from each tail to produce the Median CPI.
In December, the 12-month change in the CPI was 2 percent. But the trimmed-mean CPI estimates showed considerably higher trends, ranging from about 2-1/2 percent for the 20 percent trimmed mean, to around 3-1/2 percent for the median CPI. Each trimmed-mean estimate can give a different read on the underlying inflation trend, so which one comes closest to measuring the inflation trend that we ultimately see in the data? We find that once we trim about 8 percent off each tail of the CPI monthly price-change distribution (a 16 percent trimmed mean), we have reduced a substantial amount of the volatility in the monthly data. More extreme trims of the CPI distribution result in nearly the same stability, which means it is hard to distinguish between the relative accuracy of any trimmed-mean CPI estimator from the 16 percent trimmed-mean to the median CPI. And all of these trimmed-mean estimates appear to give a better reading of the inflation trend than the more traditional Core CPI.
How do we know that these trimmed-mean estimators don't eliminate some of the inflation "signal?" Well, we don't, and this is always a risk when appealing to any "core" inflation measures when trying to gauge underlying price pressure. Still, repeated tests in the United States and for other nations (for example, Australia) indicate that these trimmed-mean estimators seem to track the future behavior of the CPI better than either the CPI or the more traditional Core CPI.
In other words, whatever is being excluded, or trimmed, from these measures doesn't seem to be very helpful in telling us where the inflation trend is headed.