RDP 2006-10: The Performance of Trimmed Mean Measures of Underlying Inflation 1. Introduction

Headline consumer price inflation can be quite noisy, especially at high frequencies, and sometimes will not give a good measure of the trend in inflation, which is a key input for monetary policy decision making. Accordingly, many central banks focus on various measures of underlying inflation which they expect will contain less noise and may thus be helpful for meeting their ultimate objectives. Exclusion measures, which remove the effects of price changes of a fixed set of items such as food and automotive fuel, are widely used in a range of countries. Statistical measures, in particular trimmed mean measures[1], represent another approach and have received significant support from academic studies and are used by some central banks. In contrast, some other central banks downplay the use of underlying measures in their analysis of inflationary trends, focusing more on the headline consumer price index (CPI), often on the rationale that it is the ultimate goal of policy.

This study assesses the relative benefits of headline CPI, exclusion measures, and trimmed mean measures as indicators of the underlying trend in inflation.[2] We use a number of criteria, but they mostly come down to a search for measures which appear to contain a high signal-to-noise ratio. One aspect of this is that if underlying measures of inflation are successful in removing noise from CPI data, they should also be helpful in predicting near-term inflationary outcomes. Hence, an important test for various measures of underlying inflation will be the extent to which they have some predictive power for near-term inflation. Accordingly, our criteria correspond closely to Blinder's often-cited comments on his time watching month-to-month inflation as a central banker: ‘The name of the game then was distinguishing the signal from the noise, which was often difficult. The key question on my mind was typically: What part of each monthly observation on inflation is durable and what part is fleeting … To me, the durable part of the information in each monthly inflation report was the part that was useful in medium- and near-term inflation forecasting’ (Blinder 1997, p 157).

There are a number of earlier studies that assess different measures of underlying inflation, although these mostly use data for a single country.[3] However, the recent run-up in oil prices has shown that there is still a wide range of views as to the relative merits of different measures of underlying inflation – witness the different focus of some central banks on the headline CPI, others on inflation excluding the effect of fuel and food, and still others on statistical measures. Accordingly, there is significant scope for further work in this area.

We use data for four different economies – the United States, the euro area, Japan and Australia – to look for evidence on the performance of different measures of underlying inflation that is reasonably robust across these economies and thus is likely to also apply to other countries. We consider the performance of a wide range of trims, as we move from the headline CPI (0 per cent trim) to the weighted median (50 per cent trim), and also compare these against the performance of exclusion measures of underlying inflation. We also address several technical issues, including the effect of items with very large CPI weights on measures of underlying inflation. Our approach is unashamedly empirical and atheoretical, focusing not on debates over how to define underlying inflation or on the theoretical case for using trimmed mean measures[4], but rather assessing which of the measures of inflation are likely to be useful indicators for analysts both inside and outside central banks.

To preview our conclusions, we find that trimmed mean measures offer substantial advantages over headline inflation in terms of providing measures of inflation which have a higher signal-to-noise ratio. Our analysis of different degrees of trimming suggests that there is a wide range of trims that outperform exclusion measures of inflation. In addition, we find that the presence of large expenditure items has implications for the degree of trimming that will be optimal. In the case of the US CPI, we highlight the benefits of breaking up the large imputed rent of home-owners component into regional components, which yields improved measures of underlying inflation. Overall, while there are some key technical issues which must be considered when constructing trimmed mean measures of underlying inflation, we conclude that these measures are quite robust in the sense that there is a wide range of trims that perform well for all the economies we consider.


We use the term ‘trimmed mean measures’ to also encompass the weighted median, which can be viewed as the 50 per cent symmetrical trimmed mean. [1]

Some other approaches that we do not consider include: volatility weights (published by the Bank of Canada); estimates based on principal components or dynamic factor models (Cristadoro et al 2005); persistence weights (Cutler 2001); and model-based approaches (Quah and Vahey 1995). These alternatives (especially the latter two) have not been used as widely in central banks as exclusion or trimmed mean measures. [2]

For example, see Bryan and Cecchetti (1994) and Smith (2004) for the United States, Vega and Wynne (2003) for the euro area, Shiratsuka (1997) for Japan, and Kearns (1998) and Roberts (2005) for Australia. We are aware of only one other multi-country study of different measures of inflation, by Catte and Sløk (2005). [3]

Numerous other papers, including Roger (1998) and Wynne (1999), provide a discussion of the definition of core inflation; Bakhshi and Yates (1999) discuss the theoretical case for trimmed mean measures. [4]