Flexible inflation targeting has become the preferred policy among a growing number of central banks over the last decades. Due to the lag between interest rates and inflation, optimal monetary policy in this framework is essentially about forecasting inflation (Svensson and Woodford, 2003). The output gap, measuring the deviation of output from potential, has a key role in this regard. Through different transition mechanisms a positive output gap leads to inflation. For central banks aiming at a flexible inflation target, an appropriate policy response to the observed pressure in the economy will not only help stabilize inflation at a desired level, but also stabilize output (Svensson, 1997 and 2000).
If the policy reactions are going to be proper, the measure of the output gap has to be adequate. As demonstrated in this and other analysis it seldom is (see for example Orphanides and van Norden (2002) and Bernhardsen, Eitrheim, Jore and Røisland (2004). There are basically two factors making the derivation of the output gap difficult. The first concerns the estimation procedure. Since one fails to reject the hypothesis of a unit root in macroeconomic time series, the long run trend of output can no longer be treated as deterministic; see e.g. Nelson and Plosser (1982). Accordingly, the computation of potential output has to take into consideration the estimation of a stochastic trend, which greatly complicates the measuring of potential output and the output gap.
The second factor concerns the real-time nature at which central banks have to conduct monetary policy: Decisions are based on highly uncertain data, which are subjected to substantial revisions. This is especially true of the output. There are three main reasons for changes to official statistics.
• The earliest estimates are based on preliminary and incomplete information. • Changes to the base year. • The national accounts are occasionally subject to major revisions.
Real-time data is data as it was observed at each point in time, and typically categorized into different vintages describing their time of release, thus taking into account these data revision processes.
In the spirit of Orphanides and van Norden (2005), this paper examines two different methods for extracting the output gap in real-time, and evaluates their performance in forecasting Norwegian inflation. Especially, I question whether the inclusion of the output gap gives any value added in forecasting Norwegian domestic inflation compared to simple autoregressive benchmark models. The answer clearly depends on factors as model specifications, evaluation criteria, the forecasting periods and the quality of the data: The output gap models evaluated are the Hodrick-Prescott filter and the Production function method. As a benchmark forecasting model I employ a linear AR(p) model of inflation. My main forecasting model is aPhillips curve relation including the output gap. These specifications make it possible to relate inflation to real activity. I have used root mean square forecast errors (RMSFE) to assess the forecasting performance, and the forecasting period has ranged from 94q1 to 06q2. By using real-time data this paper highlights the problems and the uncertainties brought forward by the data revision processes.
To my knowledge real-time forecasting exercises of this kind has not been conducted on Norwegian data before. Bjørnland, Brubakk and Jore (2007) found that models including the output gap gave a better predictive power of inflation than models based on alternative indicators, and that they forecasted significantly better than simple benchmark models, but they did not use real-time data.
Based on real-time data estimations my findings suggests that the inclusion of the output gap makes the out-of-sample forecasts less accurate than what would have been attained if the simpler benchmark models had been used, a finding that is consistent with results reported in Orphanides and van Norden (2005). Some output gap models computed in real-time do however forecast better than the benchmark models, but the results seems to be very sensitive to the chosen forecasting period. Further I find that there are considerable differences in forecasting performance between using real-time data, and final vintage data (the final vintage in the sample has been 06q2).
The reminder of this paper is organized as follows: Section 2 describes the output gap concept, the output gap models and the real-time data sets that I have used. Sections 2-2.2 follow Bjørnland, Brubakk and Jore (2004), and Frøyland and Nymoen (2000) closely. For a more thorough exposition of the output gap, and the different methods to extract it, I refer to the cited papers. Section 2.4 illustrates clearly how the real-time issues affect the output gap estimates. Section 3 presents the forecasting methodology. Sections 4 and 5 present the results and conclusions.
I have used Matlab computer software and the Econometrics Toolbox provided by James P. LeSage for my computations. Programming codes can be made available on request.