Saturday, October 31, 2026

Why the RBA Watches the Quarterly Trimmed Mean

The RBA targets headline CPI inflation. But headline inflation is noisy. One-off price shocks push it around: the GST in 2000, free childcare and the fuel price collapse in 2020. A central bank can't set policy on that noise. It needs a measure of the underlying trend.

The ABS publishes two. The trimmed mean drops the 15% of price changes at each end of the distribution and averages the middle 70%. The weighted median takes the price change of the expenditure class that sits at the middle of the expenditure-weighted distribution.


Three tests for underlying inflation

A good underlying measure must not drift from headline over time. If the target is headline, the trend measure has to track it in the long run. Since 1987, the quarterly gap between the trimmed mean and headline averages 0.00 points. The weighted median averages 0.01. Both look unbiased in growth terms.

It must also be less volatile than headline. Both underlying measures cut the volatility of quarterly inflation by about two thirds.

And it must be persistent. This quarter's reading should tell you something about next quarter's. Headline inflation has an autocorrelation of about 0.4. Both underlying measures sit near 0.85.


Where the weighted median falls down

On noise and persistence, the weighted median currently wins by a whisker. But the price level chart shows its weakness. It wanders much further from headline. After nearly forty years, the trimmed mean index sits about 1% above headline. The weighted median sits about 1.5% above. It ran as far as 6% above in 2020, and it sat 3% to 5% above through the calm decade from 2009 to 2019. Small biases in quarterly growth compound into large gaps in the price level. For a bank that targets headline, that matters.

The reason lies in how each measure is built. The median rests on one item: whichever price change happens to land in the middle. When the distribution of price changes is lumpy, the middle can jump from one item to another, and the median lurches with it. The trimmed mean uses 70% of the basket. It throws away the outliers but keeps most of the information. That makes it harder to knock off course.

The trimmed mean also has a simpler story. It's an average with the extremes removed. Anyone can understand that.


Asymmetric trims fix a bias

Not every trimmed mean is symmetric. The Dallas Fed's trimmed mean PCE cuts 24% of expenditure weight from the bottom of the distribution and 31% from the top. The Cleveland Fed's trimmed mean CPI cuts 8% from each end.

Dallas fits its trims to the data. It chooses the proportions that best track a proxy for trend PCE inflation. In US monthly data, large price rises outnumber large price falls. A symmetric trim would sit below headline over time, so Dallas cuts more off the top. Cleveland accepts a simple rule rather than an optimised one.

Fed Chair Kevin Warsh has made the Dallas measure fashionable. He told his confirmation hearing that trimmed averages help filter out shocks and noise. But a fitted trim works only while the skew it was fitted to persists. When the distribution shifts, the calibration breaks. That is happening now. Lately the top tail has grown, so the heavier top trim throws away much of the tariff-driven price rises. Dallas Fed President Lorie Logan has said her own staff’s research cautions against reading too much into its low readings. There is a credibility cost too. Of the plausible measures, Warsh’s preferred one happens to give the lowest number.


Australia has no bias to fix

The asymmetry exists to correct bias against headline. The RBA's symmetric 15/15 trim has no bias to correct. Its quarterly gap to headline averages 0.00 points over nearly forty years. An asymmetric trim could only make that worse. A symmetric trim is less clever, but harder to game.


The monthly series reopens the question

Since November 2025, the ABS has published a complete monthly CPI. Monthly headline is now the benchmark for the inflation target. But the RBA kept its focus on the quarterly trimmed mean, and at its request the ABS will keep publishing that series for at least the first 18 monthly releases. The RBA has not yet chosen a preferred monthly measure.

The two trims are built differently. The monthly one trims monthly price changes with monthly seasonal adjustment. The quarterly one trims quarterly changes with quarterly adjustment. At first the monthly trimmed mean ran hot. Over the first five overlapping quarters, the ABS found the monthly-derived measure came in 0.22 points above the published quarterly trimmed mean. Through most of 2025, year-ended monthly trimmed mean inflation sat above the quarterly version.

Then the gap closed and reversed. Since December 2025, the quarterly trimmed mean has run slightly above the monthly one. In September 2026, the quarterly seasonally adjusted measure stood at 3.7% and the monthly at 3.6%.

Skew was the obvious suspect for the early gap. Over one month, many items don’t change price at all, and some reprice once a year in a single jump. A symmetric trim of that lumpy distribution need not land near the mean. The monthly weighted median also ran above the monthly trimmed mean through 2025. Here the Dallas logic might apply.

But a persistent skew should leave a persistent gap, and this one has flipped. That could mean the early gap came from young seasonal factors and a different set of price observations. Or the skew could return. It’s just too early to know: the monthly series only goes back to April 2024.

Dallas calibrated its trims on decades of data, and its measure is still misfiring now that the skew has shifted. Two and a half years of monthly data, with immature seasonal factors, is far too short to choose an asymmetric trim. You would be fitting noise.

For now the RBA sticks with the quarterly trimmed mean. It filters the noise without losing touch with the thing the RBA is trying to control.

1 comment:

  1. A highly informative and educational article. Many thanks

    ReplyDelete