OECD Compendium of Productivity Indicators 2018

Trend estimation method

Understanding to which extent productivity growth is driven by structural factors and affected by short-term economic fluctuations is of utmost importance for policy makers. To shed light on this distinction, one can decompose the series into a trend and a cyclical component, where the trend is meant to capture the long-term growth of the series and the cyclical component is the deviation of the series from that trend. In this publication, the method used to extract the trend component is the Hodrick-Prescott (HP) filter (Hodrick and Prescott, 1997).

The Hodrick-Prescott filter

The HP filter is the best known and most widely used method to separate the trend from the cycle (Hodrick and Prescott, 1997). The method has been first presented in a working paper in 1981 (Hodrick and Prescott, 1981). The filter is defined as the solution to the following optimisation problem:

$y_t={\tau }_t+c_t$

${min}_{\{<!--\; -->{\tau }_t\}}\{<!--\; -->\sum _{t=1}^T{\left(y_t-{\tau }_t\right)}^2+ \lambda \sum _{t=2}^{T-1}{\left[\left({\tau }_{t+1}-{\tau }_t\right)-\left({\tau }_t-{\tau }_{t-1}\right)\right]}^2 \}$

Where $y_t$is the original series, ${\tau }_t$is the trend component and $c_t$is the cyclical component. The method consists in minimising the deviation of the original series from the trend (the first term of the equation) as well as the curvature of the estimated trend (the second term). The trade-off between the two goals is governed by the smoothing parameter λ. The higher the value of λ, the smoother is the estimated trend.

For quarterly data it has been typically assumed a value of λ=1600, as recommended by Hodrick and Prescott (1997). However, there is less agreement on the value to be used when the filter is applied to other frequencies (e.g. annual, monthly). Backus and Kehoe (1992) used λ=100 for annual data, while Ravn and Uhlig (2002) propose an adjustment of the standard value of 1600 that consists of multiplying that value by the fourth power of the frequency of observations relative to quarterly data. The latter results in a value of λ equal to 6.25 (=1600*(1/4)⁴) for annual data.1

The HP-filter can be interpreted in the frequency domain. In this formulation the λ parameter can be associated with the cut-off frequency of the filter – the frequency at which it halves the impact of the original cyclical component. It can be shown that the Ravn-Uhlig rule for selecting the value of λ corresponds to a cut-off frequency of approximately 10 years, assuming annual data (Maravall and Del Río 2001). Nonetheless, Nilsson and Gyomai (2011) point out that the HP-filter has strong leakages (i.e. letting cyclical components from the stop band appear in the filtered series), and this feature may affect the choice of the filter parameter depending on the goal of the study and sensitivity to filter leakage.

In this publication, the target frequency for trend estimation was no different than in the above studies (10 years and beyond). However an additional objective is to minimize the leakage from shorter business-cycle frequencies into the estimated trend. Accordingly, the value of the smoothing parameter selected here is λ=54.12. This value has been determined by calibrating the Hodrick-Prescott filter in such a way that the frequency response at 9.5 years is equal to 0.10. This means that with λ=54.12, cycles with a wavelength lower than 9.5 years would be attenuated by 90% or more.

In comparison with other ideal filters, the trend estimated with the HP filter is more sensitive to transitory shocks or short-term fluctuations at the end of the sample period. This results in a sub-optimal performance of the HP filter at the endpoints of the series (Baxter and King, 1999). In view of this property, in order to lessen revisions of the published estimates, trend series are not published for the first two years and the last two years for which data on the original series are available. Even though, the choice of the HP filter is based on its interpretability and widespread use in the literature.

Further reading

Backus D. and P. Kehoe (1992), “International evidence on the historical properties of business cycles”, The American Economic Review, Vol. 82, No. 4.

Baxter and King (1999), “Measuring business cycles: approximate band-pass filters for economic time series”, The Review of Economics and Statistics, Vol. 81, No. 4.

Hodrick R. and E. Prescott (1981), “Postwar U.S. business cycles: an empirical investigation”, Carnegie Mellon University, Discussion paper No. 451.

Hodrick R. and E. Prescott (1997), “Postwar U.S. business cycles: an empirical investigation”, Journal of Money, Credit and Banking, Vol. 29, No. 1.

Maraval, A. and A. Del Río (2001), “Time aggregation and the Hodrick- Prescott filter”, Banco de España, Servicios de Estudios, Documento the trabajo No. 0108.

Nilsson, R. and G. Gyomai (2011), “Cycle Extraction: A Comparison of the Phase-Average Trend Method, the Hodrick-Prescott and Christiano-Fitzgerald Filters”, OECD Statistics Working Papers, No. 2011/04, OECD Publishing, Paris, http://dx.doi.org/10.1787/5kg9srt7f8g0-en.

Ravn M. and H. Uhlig (2002), “On adjusting the Hodrick-Prescott filter for the frequency of observations”, The Review of Economics and Statistics, Vol. 84, No. 2.

Note

← 1. The frequency of observations relative to quarterly data is 1/4 for annual data and of 3 for monthly data.