An answer, from the latest edition of the Federal Reserve Bank of Cleveland's Economic Trends:
Identifying changes in the inflation trend is generally only possible after long periods of time have passed. Moreover, methods to measure the underlying inflation pattern in the data, such as long-run averages, can reveal a shift in the inflation trend only well after that change has occurred.
To improve the inflation signal in the price data, economists have often appealed to so-called core inflation measures, like the CPI excluding food and energy items—goods notorious for causing transitory fluctuations in the aggregate price data. A more recent approach is the use of trimmed-mean estimates that systematically strip out the more extreme—and presumably most transitory—price changes. These measures have been shown to substantially reduce short-run variation in the inflation estimates and, hopefully, give policymakers a quicker read on shifts in the inflation trend. Indeed, these estimates have predicted the long-term growth rate of the CPI better than either the CPI or the more traditional CPI excluding food and energy. For example, since 1990, monthly changes in the median CPI and the 16% trimmed-mean CPI have been about twice as effective as changes in the overall CPI for predicting the longer-term CPI inflation trend (that is, the 36-month annualized percent change).
The following picture is meant to illustrate the superiority of core measures of inflation in predicting headline inflation over a 3-year horizon:
The way to read this graph is as follows: The points corresponding to the number 1 on the horizontal axis answer the question "What is the root-mean-squared error of a prediction of inflation over the next 36 months, where those predictions are based on one month's observations on the CPI itself (the red line), the CPI excluding food and energy components (the blue line), the median CPI (the yellow line), and a core measure with the most extreme 16% of prices stripped out (the purple line)?" (If you are not familiar with statistics, the lower the root mean-squared error, the better the prediction.) The points associated with the number 2 on the horizontal axis correspond to the question "which measure is a better predictor of longer-run inflation if we use 2 months of data?" The points corresponding to the number 3 answer the question "which measure is a better predictor of longer-run inflation if we use 3 months of data", and so on.
If what we are interested in is getting an accurate guess about what inflation is going to be over the next several years, then this exercise indicates that (a) it is better to focus on inflation over the past 9-12 months than to extrapolate on the basis of just one or two months; and (b) it is better to focus on core inflation than headline inflation.
Something to keep in mind when the March CPI numbers roll out next Wednesday.