RDP 2009-06: Inflation Volatility and Forecast Accuracy 1. Introduction

Over the twenty years to 2008, the level and volatility of inflation has declined across the world (Table 1). Average CPI inflation across the major countries fell from 6.0 per cent over 1977–1992 to 2.0 per cent over 1993–2008, while the unconditional standard deviation fell from 3.7 per cent to 1.6 per cent over the same period. While these trends are common to all countries, the extent of change has varied across countries consistent with the tendency for convergence of both the level and volatility of inflation.^[1]

There is a literature examining the role of monetary regimes in explaining these changes, with one particular focus on differences between inflation-targeting (IT) and non-IT regimes (see, for example, Bernanke et al 2001). Others, such as Geraats (2002), Chortareas, Stasavage and Sterne (2001) and Demertzis and Hughes Hallett (2007), study how the precise nature of the policy framework, such as the degree of central bank transparency, is related to the volatility and level of inflation. Many of these studies rely on simple measures of inflation behaviour, such as unconditional means and variances, and are usually based on headline measures of inflation. However, such measures can be overly influenced by very temporary movements in inflation. This means that the sample period for the analysis can have an important influence on the results. More importantly, these temporary effects may have little if anything to do with differences in policy frameworks and much more to do with different structural features of the economy, such as its size or openness to trade. One alternative is to focus on underlying or core measures of inflation. However, there is no widespread agreement on the best way to do this, and comparable measures across a wide range of countries are not readily available. Another alternative is to use a statistical model to try to separate headline inflation into persistent and temporary components.

This paper adopts this latter approach, examining the inflation process in five IT countries (Australia, Canada, New Zealand, Sweden and the United Kingdom) and five non-IT countries (Austria, France, Germany, Japan and the United States) using a statistical model introduced by Stock and Watson (2007). The Stock and Watson approach decomposes inflation into permanent and transitory components, the variabilities of which are allowed to change over time. Using measures based on this unobserved components stochastic volatility (UC-SV) model, we find little support for sharp distinctions between countries in terms of the level and volatility of the permanent component of inflation.^[2]

A related approach is to examine the forecastability of inflation.^[3] An effective and credible monetary policy regime, other things equal, will help to keep inflation anchored closely around a low and constant mean. By itself, this implies that inflation should be easier to forecast – that is, forecast errors will tend to be small.

To examine the forecastability of inflation, we used a modified version of the Stock and Watson model (M-UC-SV). Following a suggestion of Pagan (2008), we model the time-varying volatilities as autoregressive processes of order one (AR(1)), so that they have finite second moments. And instead of assuming that the permanent level of inflation follows a random walk, we use a mean-reverting AR(1), with a freely estimated degree of persistence. These assumptions imply that inflation is a stationary process that can be decomposed into temporary and persistent (but not permanent) components, consistent with the notion that monetary policy can influence inflation and provide a nominal anchor.

In brief, we find that inflation forecastability improved over time across our sample of selected IT and non-IT countries, both in absolute terms and relative to naïve forecasts. Across countries, the out-of-sample forecast error of the M-UC-SV model tends to be somewhat smaller than that of the original UC-SV model. Furthermore, it seems that the improvement in inflation forecastability was more pronounced in IT countries than in non-IT countries.

The rest of the paper is organised as follows. Section 2 describes the UC-SV model and its modified version M-UC-SV together with a description of the data used in the analysis. Section 3 presents the within-sample results on the inflation process based on the UC-SV model. Section 4 presents results on the forecastability of inflation where the focus is on the M-UC-SV model. Section 5 concludes.

Footnotes

The decline in inflation volatility has not come about because central banks were willing to tolerate higher output volatility. In fact, at least over the period up to 2008, output volatility has tended to decline (see Table F1; Stock and Watson 2005 and Kent, Smith and Holloway 2005). It may be that economies have faced a more benign inflation-output volatility trade-off over this period and/or that policy has played some role – see Cecchetti, Flores-Lagunes and Krause (2006), for example. [1]

Related work has documented the quantitative effects of inflation targeting. Kuttner and Posen (2001) document that inflation targeting reduces the persistence of inflation. Benati (2008) concludes that inflation is highly persistent in policy regimes that lack a well-defined nominal anchor. Pivetta and Reis (2007) find that inflation persistence has been high and approximately unchanged in the United States since 1965. Cogley, Primiceri and Sargent (2008) argue that this finding can be viewed simply as a manifestation of shifts in average (or the target for) inflation. They conclude that inflation persistence has decreased since the 1980s. [2]

Earlier literature has looked at survey-based inflation expectations. Levin, Natalucci and Piger (2004) find that inflation targeting is effective in anchoring inflation expectations. Johnson (2002) finds that the level of expected inflation in targeting countries falls after the announcement of targets. However, neither the variability of surveyed inflation expectations nor the average absolute survey-based forecast error fall after the announcement of targets. [3]