Trimmed-mean inflation measure and its sibling median inflation gauge have both gotten unprecedented attention in recent months. It is all owing to new Federal Reserve chairman Kevin Warsh.
In his Senate Banking committee confirmation hearing in April, he said: “We used to use core PCE measures, so we’d exclude food and energy, because it was sort of a rough swag as to what was going on. We don’t have to do a rough swag anymore. What I’m most interested in is, what’s the underlying inflation rate.”
“The measures I prefer are looking at things that are called trimmed averages, where we take out all of the tail risks, all of the one off items, and we ask ourselves whether the generalized change in prices is having second order effects on the economy,” Warsh added. “Trimmed averages” is another way to say “trimmed mean”.
Trimmed-mean and median inflation measures both try to isolate the so-called “underlying” price trend by sorting every basket component by its monthly price change and looking past the extremes. Median inflation keeps only the single component sitting at the 50th percentile; trimmed mean excludes a fixed slice from both tails and averages what remains to extract the signal.
They are both advanced versions of the “core” inflation measure we generally adopt in major economies including the US, which excludes certain components that are deemed most volatile. Food and energy prices are the most common example; sometimes tobacco and alcohol prices are also excluded.
“Simple core” measures assume what is the most volatile. Trimmed-mean and median versions look at the data each month and cut out the ones that really fluctuated the most, aiming to improve the noise filtering efficiency.
This sounds like an improvement in how we measure inflation, right? Why wait for some Warsh Task Forces to evaluate it? As a matter of fact, the Fed system also has these measures available. The Dallas Fed maintains a dataset for trimmed-mean PCE inflation while the Cleveland Fed keeps track of trimmed-mean and median CPI as well as median PCE inflation. Why wait?
Been There, Trimmed That
Well, north of the border, the Bank of Canada has actually adopted trimmed-mean and median inflation gauges as its core inflation indicators since 2016. Its experiences may serve as a cautionary tale for the Fed.
BoC in 2016 adopted CPI-trim (i.e. trimmed mean, with a 20% trim at each tail) and CPI-median as its core inflation measures. The measures this year face their second five-year monetary policy framework review and they are facing some public doubts.
Late last month, BoC published the comments it received from the monetary policy framework consultation. Regarding the core inflation measures, the Bank said: “Private sector economists and think tanks… suggested the Bank should focus more on total CPI in public communications, since it is the target and because the public is skeptical of measures of core inflation.”
On top of that, “[m]ost participants in the community conversations found core inflation to be a challenging concept to grasp, despite being provided with various explanations,” according to the comments.
Using these “advanced” techniques didn’t help the public understand the cost-of-living situation any better—if anything, they are branded as confusing or challenging [footnote 11]. It makes one wonder: are they really better core inflation measurements?
This is why in this article we review some of the complaints heard in Canada in the past decade so we can get a sense of the downsides of trimmed means.
To begin with, let me first introduce two of the most prominent complainants against the trimmed mean: Matthieu Arseneau and Alexandra Ducharme, both economists at the National Bank of Canada (NBC). They wrote a series of research notes from 2023 to 2025 pointing out the downsides of using CPI-trim and CPI-median as the official core inflation measures.
Shelter’s Shadow Influence
Let’s start with a technical complaint—some of the heavily weighted components can skew the indicator and if the compilation of the inflation of these components is somehow biased, it can have a substantial impact on trimmed mean’s ability to estimate underlying inflation. One notable example is housing costs.
In March 2024, Bank of Canada Governor Tiff Macklem said:
“If you look beyond shelter, we are seeing that underlying inflationary pressures persist. And one way to look at that is, if you look at our preferred measures of core (CPI-trim, CPI-median), those exclude the things that are going up the most and the things that are going down the most. Most of the shelter components are in the things that are going up the most, so they’re excluded from those core measures. Those core measures are still running over 3%.”
Arseneau and Ducharme disagreed with Macklem in their research note, arguing that even though CPI-trim had excluded housing costs from the calculation, they can still have an outsized impact on the trimmed mean.
“Given their imposing weight (11% of the total basket), these categories act as magnets for the core measures of inflation. In other words, they occupy the space of the components at the top of the distribution range, and thus indirectly contribute to the rise in inflation calculated by these measures,” the two NBC economists explained.
They accompanied the analysis with the following graphic, illustrating how the housing costs can pull CPI-trim and CPI-median higher even when they are excluded from the calculation:
Nonetheless, we have to remind you of a Canada-specific context: Statistics Canada calculates housing costs differently from the US’s BLS. In the US, owned accommodation expenses are computed as the opportunity cost of living in the house instead of renting it out, i.e. the so-called owner’s equivalent rent (OER), an “imaginary” expense that doesn’t really happen.
In Canada, however, the approach is different. StatCan tries to capture the out-of-pocket expenses of owning and living in the house, so one of the components it uses is mortgage interest cost (MIC). The problem is that MIC is, after all, an interest expense. As the BoC controls short term interest rates, it indirectly influences MIC through its monetary policy. This results in an inherent conflict.
MIC, as stated above, has a non-negligible weight in the CPI calculation, meaning that BoC can potentially exert influence on the inflation gauge it targets. What Arseneau and Ducharme argued was that this problem is worse when the BoC was changing its benchmark interest rate in a rapid sequence like it did in 2022 and 2023.
Here the NBC researchers compared the core measures with a hypothetical scenario in which the MIC and rent stayed at their 2009–2019 average levels, showing that CPI-trim and CPI-median may be “overestimated” by the increase in MIC and rent.
Does this have any relevancy to the US’s situation, though? While the Fed can’t directly influence OER, the opportunity cost of living in one’s own home still counts as a very substantial part of US CPI calculations. Any unusual behavior exhibited in this component can skew the overall CPI.
Cleveland Fed tried to tackle this problem of an outsized component biasing the calculation of trimmed-mean CPI. In their July 2007 trimmed-mean methodology adjustment, they broke down OER—which had a 23.83% weight back then—into four regions, so each region had a weight between 4.83%–7.23% (also December 2006 figures). The change “improve[d] the ability of the trimmed-mean measures to predict the future CPI trend, as greater percentages of the extreme components are trimmed from the data,” according to their backtest with January 1983 to June 2004 data against 36-month forward CPI inflation.
Still, one of the complaints we often hear about OER is that it is a lagged indicator of the “true” housing expenses. That is because the way BLS measures OER and rent is based not only on the rents of the newly signed leases of the month, but of all existing rental values surveyed. Hence, the growth rate of this component changes much slower than the new tenant rents in the market.
The gradual decline of OER was often cited as one of the reasons the disinflation process flatlined in 2024 and early 2025. This also affected trimmed-mean CPI’s disinflationary trend. Using a similar method to Arseneau and Ducharme, we created a hypothetical scenario in which OER stayed at the level of the decade from 2009 to 2019.
The OER stable trimmed-mean CPI is, as expected, lower than the Cleveland Fed version. The focus, though, should be on how long it has taken for the gap to close since 2022.
Another scenario is to calculate trimmed-mean CPI using the BLS’s experimental new tenant rent index instead of OER.
The new tenant rent index only counts the rent on newly signed leases. As shown in the red line, incorporating it in trimmed mean would make the measure peak at a much higher rate in 2022. But, importantly, it also fell much faster in 2023, instead of the slow burn we saw in the original version (blue dashed line).
While trimmed mean can exclude the volatile component in CPI calculations, it can’t overcome the lagging nature of OER and still suffers from the slow disinflation process, just like the overall inflation rate.
The Trouble of Swinging Skewness
Another technical concern is that the changing skewness of the inflation distribution can affect the accuracy of trimmed-mean measures. Lorie Logan, president of the Dallas Fed, pointed out this problem in a speech in June: “A change in the mix of price increases and decreases is causing the trimmed mean to drop too many price increases. That can pull the trimmed mean below the underlying trend in inflation.” She cited a study written by three Dallas Fed economists in April which looked into the regional Fed’s own trimmed-mean PCE measure.
The Dallas Fed applies an “asymmetric trim” to the inflation gauge, excluding the top 31% and bottom 24% components—i.e., more of the fastest rising prices are trimmed out compared to the slowest price increases. This arrangement is to handle the “negative skewness” found in the 1977–2009 US PCE inflation data—the bottom 10 percent of price changes are much further below the median of the distribution than the top 10 percent are above the median. If they trim the tails equally, like in BoC’s method, it can lead to an overestimation problem because, in oversimplified terms, the highly negative numbers are cut out of the calculation.
The problem Logan referred to is that in recent years the price change distribution switched to become much less negatively skewed, and even positively skewed during Covid and recently.
In such cases, the asymmetric trim resulted in an underestimation of the inflationary trend. This is one of the leading explanations of why the Dallas Fed trimmed-mean PCE was slow to pick up the Covid inflation surge.
So, this is a problem with the asymmetric trim method, right? Just use an equal trim and we should solve the problem? If we look at trimmed-mean CPI by the Cleveland Fed, which implements an equal trim of 8% on each side, it does show a better performance in catching the Covid inflation spike.
Similar problems happened in Canada, even though the BoC adopted the equal trim method. In its April 2021 Monetary Policy Report, the Bank cast doubt on the two core measures’ performance:
“Early in the pandemic, both CPI-median and CPI-trim excluded components with large, sudden price declines. These declines were felt in the most heavily affected sectors of the economy. By excluding these components, the measures had been putting more weight on components that were not affected (e.g., property taxes). However, when the prices of hard-hit components stopped falling and started to recover gradually, these measures tended to include the associated increases. This resulted in an asymmetry. Because their composition is based on monthly price movements, these measures do not account for the fact that the price increases have been from particularly low levels. Consequently, CPI-median and CPI-trim have likely been overestimating underlying inflation.”
As the following graph shows, BoC expressed concern that the trimmed-mean and median measures failed to handle the rapid price changes in the 2020–2021 pandemic period.
Then in March 2024, Desjardins Macro Strategists Royce Mendes and Tiago Figueiredo pointed out that:
“The distribution of Canadian consumer price changes is heavily skewed at the moment, suggesting bias in CPI‑median and CPI‑trim… A naive median or trimmed mean measure, therefore, completely ignores the depth of price declines across a number of sectors in the fat left tail of the distribution.”
They suggested that BoC adopts an asymmetric trim, like the one used by the Dallas Fed’s trimmed-mean PCE, to better handle “overestimation” introduced by the negative skewness in the data.
But is that a wise strategy? Below is a Canada CPI skewness time series. We can see that while in both H1 2021 and H1 2024 the readings dropped into negative territory, the figures fluctuated quite a lot, especially around the pandemic era.
This is the same problem Logan cited. When the inflation rate skewness changes rapidly, the reading can be “misleading” in the short term. However, unless we can make sure the fat tail only falls on one particular side, like the US’s case between 1980 and 2010, it is hard to be sure whether a permanent asymmetric trim is appropriate. But without appropriate adjustments, trimmed mean can miss out some of the signal in the overall inflation trend, like its failure to predict the Covid inflation surge in the US.
An Invitation to Cherry Pick
Another common complaint is related to the Canadian central bank’s policy implementation—it introduced too many core inflation measures in its communications.
Here is an excellent graphic NBC’s Arseneau and Ducharme included in their November 2025 note to illustrate the problem:
Before switching to CPI-trim and CPI-median, BoC used some variants of simple core: From 1991 to 2001, it excluded food, energy and indirect taxes (named CPI-XFET by BoC convention); from 2001 to 2016, it became CPIX, excluding the eight most historically volatile basket components, along with indirect taxes.
While the Bank switched to the two advanced core measures—along with CPI-common, which uses a statistical model to find out the underlying inflation trend by removing the volatile components—it didn’t stop the central bank from picking those old measuring tools up when it saw fit. This led to an apparent cherry-picking problem.
This may sound relevant to the current Warsh situation, as analysts wonder whether Warsh’s preference for the “trimmed average” inflation measure was that it happened to be the one with the lowest inflation reading at the moment.
Introducing more core measures by what is ripe for the season, even with the best intentions, can be “confusing” for the public. In the BoC’s community outreach, participants expressed that “[b]ringing core inflation into the discussions created confusion about the target and what the Bank assesses when it makes its policy decisions.”
As then Deputy Governor of BoC Rhys R. Mendes said in a speech in October last year: “We’ve said these preferred measures are ‘an operational guide to help the Bank achieve its inflation target.’ At times, this language may have led markets to place more emphasis on the preferred core measures than we do,” adding that the central bank “don’t want Canadians or financial markets to become overly focused on a single indicator.”
It can be an example that central banks generally prefer to have more inflation measurement tools in their toolkit, giving them maximum flexibility.
The “confusion” faced by the general public, in its view, is more of a communication efficiency issue, which Mendes suggested solving by “publishing an interactive dashboard next year (2026) to house our broad array of inflation indicators.” This is one of the challenges switching core inflation measures can introduce.
Nonetheless, BoC hasn’t wasted the decade of CPI-trim and CPI-median implementation. In the last couple of years, the Canadian central bank has often used the percentage of CPI components rising above 3% as a way to communicate whether they think inflation is under control or not. This is a relatively easy-to-understand concept, compared to the rationale of excluding certain inflation components from calculations.
This so-called inflation breadth can be the one positive recent lesson the Fed and the Warsh Task Force can learn from.
On July 14, during his first testimony before the House Financial Services Committee, Warsh pushed back on the idea that he prefers trimmed-mean measures of inflation, saying, “if I had a preferred measure, I wouldn’t call for a task force to go back to first principles.”
As we have illustrated in this article, this “open-mindedness” is indeed the best thing to do as trimmed mean may not be as perfect as some think.
- Footnote 1: Though, Canadians also explicitly hate simple core measures, as “[community conversations p]articipants viewed food and energy as essential goods, and they disagreed with excluding these from measures of core inflation.” ↩︎
Trimmed-mean and median inflation measures both try to isolate the so-called “underlying” price trend by looking past the extremes.
🫵 Median keeps only the single component sitting at the 50th percentile;
✂ Trimmed mean excludes a fixed slice from both tails and averages what remains
BoC in 2016 adopted CPI-trim (i.e. trimmed mean, with a 20% trim at each tail) and CPI-median as its core inflation measures.
Late last month, BoC published the comments it received from the monetary policy framework consultation… they got comments like these
Using these “advanced” techniques didn’t help the public understand the cost-of-living situation any better
Are they really better core inflation measurements?
In this article we review some of the complaints heard in Canada in the past decade so we can get a sense of the downsides
Also, an important point is … one part of the article is about “Cherry Picking”
So, I can use the photo I took going Cherry Picking at St Catherine ~~~
😃
I’ve lost the thread here w/ Warsh & trimmed mean, what’s the perceived problem with core-PCE again?
too simple? actually here St Louis Fed as a good piece on this today