Assessing French inflation persistence with impulse saturation break tests and automatic general-to-specific modelling

Abstract

This article has three different motivations. Firstly, we wish to contribute to the debate on whether French inflation has been persistent since the mid-eighties. Empirical evidence in this domain has been mixed. We use the standard method of testing for breaks in the mean of the inflation series to conclude whether possible unit root findings are the result of neglected breaks. Then, we build standard autoregressive representations of inflation, using an automatic general-to-specific approach. We conclude against inflation persistence in the sample period, and the point estimates of persistence we obtain are several percentage points below those achieved with other break tests and model selection methods. Moreover, our final model is congruent. Secondly, we provide the first empirical application of the new impulse saturation break test. The resulting estimates of the break dates are in line with other literature findings and have a sound economic meaning, confirming the good performance the test had revealed in theoretical and simulation studies. Finally, we also illustrate the shortcomings of the Bai–Perron test when applied to a small sample with high serial correlation. Indeed, we show the Bai–Perron break dates’ estimates would not allow us to build a congruent autoregressive representation of inflation.

Acknowledgements

This article was written while the first author was a graduate research visitor at the European Central Bank. The first author's doctoral supervisor, David F. Hendry provided very valuable comments. Useful insights were gained from many conversations with Ignazio Angeloni, Benoit Mojon, Laurent Bilke, Kristoffer Nimark, Andrew Levin, Jeremy Piger, Andreas Bayer and Victor Gaspar. The authors are especially thankful to to Soren Johansen, Bent Nielsen, Adrian Pagan, Jurgen Doornik and Gunnar Bardsen who took the time to read and comment on the manuscript on several occasions. Comments from the Editor and three anonymous referees were] most useful. The usual disclaimer applies. Financial support from the Fundação para a Ciência e a Tecnologia, Lisboa, is gratefully acknowledged by the first and second authors, respectively.

Notes

¹All data used in Levin and Piger (2004) were collected from the OECD statistical compendium. We are grateful to Jeremy Piger for having sent us this data.

²The period of insignificant dummies between 1989Q1 and 1990Q1 is of insufficient length, according to the aforementioned criterion, for a break point and (hence) a change in regime to be considered. Furthermore, the indicator for 1989Q3 had a p-value of 0.04, yielding one significant dummy in that period. 1991Q2 is an even more clear example of a case where only one indicator deviates from the mainstream of the period in terms of p-values. It is therefore the case that this cannot be regarded as a change in regime. These examples highlight the need to define a minimum regime length.

³Not reproduced here for space considerations, but available upon request.

⁴These insignificant indicators can be easily linked to events in French economic policy history, as accounted for in Bilke (2004b). According to Bilke (2004b), one-time changes in the VAT rates are good candidates to explain such insignificant dummies: e.g. the insignificant dummy for 1989Q1 is probably due to the simultaneous decreases in the VAT high rate and in the VAT low rate that took place in January 1989; whilst the decrease in the VAT rate for specific items in the CPI in January 1990 could explain the lack of significance of the 1990Q1 dummy.

⁵This is clearly greater than ε = 0.05, the value suggested in Bai and Perron (1998). Our choice reflects precisely the serial correlation in the data: allowing for regimes with more observations in order to improve the estimate of the variance.

⁶Bai and Perron (2003a) recommend the default use of M = 5. However, given the small sample size we are dealing with, this would force us to reduce the trimming factor, obtaining more imprecise nonparametric variance estimates.

⁷The Ox codes used in this article were written with the invaluable help of Jack Luchetti and are available upon request.

⁸PcGets is an Ox (Doornik, 2001) package, designed to implement automatic GETS model selection (see Hendry and Krolzig, 2001).

⁹In Table 7, AR stands for the Breusch-Godfrey autocorrelation test (Breusch, 1978; Godfrey, 1978), ARCH stands for Engle's (1982) test and RESET stands for Ramsey's (1969) test for mis-specification errors. The normality test used is the one suggested by Hansen and Doornik (1994), whilst White's (1980) test is used to check homoscedasticity.